segresnet-mlx 0.3.0__tar.gz → 0.3.1__tar.gz

{segresnet_mlx-0.3.0 → segresnet_mlx-0.3.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: segresnet-mlx
-Version: 0.3.0
+Version: 0.3.1
 Summary: Native MLX SegResNet with NumPy sliding-window logit inference
 Project-URL: Homepage, https://huggingface.co/MedMLX/segresnet-renal-cect-mlx
 Author: Joseph Sandoval, Alfredo Sandoval Jr.
@@ -26,7 +26,7 @@ Description-Content-Type: text/markdown
 
 Segment brain tumors, renal structures and whole-body CT anatomy with MONAI SegResNet models in native MLX.
 
-[BraTS weights](https://huggingface.co/MedMLX/segresnet-brats-mlx) · [Renal weights](https://huggingface.co/MedMLX/segresnet-renal-cect-mlx)
+[Docs](https://github.com/MedMLX/segresnet-mlx/blob/main/docs/technical.md) · [BraTS weights](https://huggingface.co/MedMLX/segresnet-brats-mlx) · [Renal weights](https://huggingface.co/MedMLX/segresnet-renal-cect-mlx) · [MedMLX](https://github.com/MedMLX)
 
 | Input | Output |
 | :--- | :--- |
@@ -34,14 +34,19 @@ Segment brain tumors, renal structures and whole-body CT anatomy with MONAI SegR
 
 **Apple Silicon · Metal · Python 3.13 / 3.14** · [uv](https://docs.astral.sh/uv/)
 
-## Quick start
+## Install
 
 ```bash
 uv venv --python 3.13
 source .venv/bin/activate
-uv pip install "segresnet-mlx>=0.3.0"
+uv pip install "segresnet-mlx @ git+https://github.com/MedMLX/segresnet-mlx.git@f60afe77518f5e617145cc4ccfa8eaa59e9eeb3e"
 ```
 
+This source revision validates NIfTI units, voxel values and geometry before
+inference; PyPI 0.3.0 predates those checks.
+
+## Run
+
 ```python
 from segresnet_mlx.medmlx import brats
 
@@ -70,6 +75,6 @@ weights report `null` unless the folder carries a MedMLX download revision.
 
 ## License
 
-Apache-2.0; license and third-party notices are included in the distribution. Model weights retain their publisher terms.
+[Apache-2.0](https://github.com/MedMLX/segresnet-mlx/blob/main/LICENSE). See [third-party notices](https://github.com/MedMLX/segresnet-mlx/blob/main/THIRD_PARTY_NOTICES.md). Model weights retain their publisher terms.
 
 For research and education. Outputs require human review.

{segresnet_mlx-0.3.0 → segresnet_mlx-0.3.1}/README.md

@@ -2,7 +2,7 @@
 
 Segment brain tumors, renal structures and whole-body CT anatomy with MONAI SegResNet models in native MLX.
 
-[BraTS weights](https://huggingface.co/MedMLX/segresnet-brats-mlx) · [Renal weights](https://huggingface.co/MedMLX/segresnet-renal-cect-mlx)
+[Docs](https://github.com/MedMLX/segresnet-mlx/blob/main/docs/technical.md) · [BraTS weights](https://huggingface.co/MedMLX/segresnet-brats-mlx) · [Renal weights](https://huggingface.co/MedMLX/segresnet-renal-cect-mlx) · [MedMLX](https://github.com/MedMLX)
 
 | Input | Output |
 | :--- | :--- |
@@ -10,14 +10,19 @@ Segment brain tumors, renal structures and whole-body CT anatomy with MONAI SegR
 
 **Apple Silicon · Metal · Python 3.13 / 3.14** · [uv](https://docs.astral.sh/uv/)
 
-## Quick start
+## Install
 
 ```bash
 uv venv --python 3.13
 source .venv/bin/activate
-uv pip install "segresnet-mlx>=0.3.0"
+uv pip install "segresnet-mlx @ git+https://github.com/MedMLX/segresnet-mlx.git@f60afe77518f5e617145cc4ccfa8eaa59e9eeb3e"
 ```
 
+This source revision validates NIfTI units, voxel values and geometry before
+inference; PyPI 0.3.0 predates those checks.
+
+## Run
+
 ```python
 from segresnet_mlx.medmlx import brats
 
@@ -46,6 +51,6 @@ weights report `null` unless the folder carries a MedMLX download revision.
 
 ## License
 
-Apache-2.0; license and third-party notices are included in the distribution. Model weights retain their publisher terms.
+[Apache-2.0](https://github.com/MedMLX/segresnet-mlx/blob/main/LICENSE). See [third-party notices](https://github.com/MedMLX/segresnet-mlx/blob/main/THIRD_PARTY_NOTICES.md). Model weights retain their publisher terms.
 
 For research and education. Outputs require human review.

{segresnet_mlx-0.3.0 → segresnet_mlx-0.3.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "segresnet-mlx"
-version = "0.3.0"
+version = "0.3.1"
 description = "Native MLX SegResNet with NumPy sliding-window logit inference"
 readme = "README.md"
 requires-python = ">=3.13,<3.15"

{segresnet_mlx-0.3.0 → segresnet_mlx-0.3.1}/src/segresnet_mlx/__init__.py

@@ -19,7 +19,7 @@ if TYPE_CHECKING:
     from segresnet_mlx.weights import load_model as load_model
     from segresnet_mlx.weights import required_shapes as required_shapes
 
-__version__ = "0.3.0"
+__version__ = "0.3.1"
 
 _EXPORTS = {
     "BUNDLE_CONFIGS": "segresnet_mlx.predictor",

{segresnet_mlx-0.3.0 → segresnet_mlx-0.3.1}/src/segresnet_mlx/_nifti.py

@@ -32,10 +32,22 @@ class NiftiHeader(Protocol):
         """Set voxel spacing in the header."""
         ...
 
+    def get_xyzt_units(self) -> tuple[str, str]:
+        """Return the NIfTI spatial and temporal unit names."""
+        ...
+
+    def set_xyzt_units(self, xyz: str) -> None:
+        """Set the NIfTI spatial unit name."""
+        ...
+
     def __getitem__(self, key: str) -> HostArray:
         """Return a named NIfTI header field."""
         ...
 
+    def __setitem__(self, key: str, value: int) -> None:
+        """Set an integer-valued NIfTI header field."""
+        ...
+
 
 class NiftiImage(Protocol):
     shape: tuple[int, ...]

{segresnet_mlx-0.3.0 → segresnet_mlx-0.3.1}/src/segresnet_mlx/_records.py

@@ -60,6 +60,7 @@ class Details(TypedDict):
     weights_revision: NotRequired[str | None]
     bundle: str
     arithmetic_profile: ArithmeticProfile
+    streaming: NotRequired[bool]
     labels: dict[str, str]
     labels_present: list[int]
     input_shape: list[int]

{segresnet_mlx-0.3.0 → segresnet_mlx-0.3.1}/src/segresnet_mlx/_types.py

@@ -6,14 +6,19 @@ from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
     from collections.abc import Callable, Mapping, Sequence
-    from typing import Literal, Protocol
+    from typing import Literal, Protocol, TypedDict
 
     import mlx.core as mx
     from medmlx_core.typing import MlxRuntime
+    from medmlx_core.volume_workspace import VolumeWorkspace
 
     from segresnet_mlx._arrays import Float32Array, HostArray
 
     type Array = mx.array
+
+    class WorkspaceOptions(TypedDict, total=False):
+        workspace: VolumeWorkspace
+
     type ArrayValue = bool | int | float | complex | list[ArrayValue] | tuple[ArrayValue, ...]
 
     class ArrayConstructor(Protocol):
@@ -51,6 +56,7 @@ if TYPE_CHECKING:
             with_coord: bool = False,
             predictor_input: Literal["numpy"],
             preserve_window_strides: bool = False,
+            workspace: VolumeWorkspace | None = None,
         ) -> Float32Array:
             """Blend predictor outputs from overlapping windows into host logits."""
             ...

{segresnet_mlx-0.3.0 → segresnet_mlx-0.3.1}/src/segresnet_mlx/medmlx.py

@@ -6,6 +6,7 @@ import os
 import tempfile
 import time
 from collections.abc import Mapping
+from contextlib import nullcontext
 from dataclasses import dataclass
 from pathlib import Path
 from typing import TYPE_CHECKING
@@ -14,6 +15,8 @@ from segresnet_mlx._arrays import Float32Array, Float64Array, MaskArray
 from segresnet_mlx._records import RunResult, Runtime
 
 if TYPE_CHECKING:
+    from medmlx_core.volume_workspace import VolumeWorkspace
+
     from segresnet_mlx.network import SegResNet
 
 import numpy as np
@@ -300,6 +303,8 @@ def run(
     image_path: str | None = None,
     overwrite: bool = False,
     arithmetic_profile: ArithmeticProfile = "torch-order-fp32",
+    streaming: bool = False,
+    scratch_dir: str | None = None,
 ) -> RunResult:
     """Segment one case with a MONAI SegResNet bundle and write the label map.
 
@@ -353,6 +358,10 @@ def run(
         overwrite: Replace an existing output mask when true; defaults to false.
         arithmetic_profile: ``torch-order-fp32`` (default), ``compensated-fp32``,
             or ``compensated-fp48``; the last supports only the renal bundle.
+        streaming: Use bounded blending intermediates and temporary mapped logits.
+            The final label volume remains resident; defaults to False.
+        scratch_dir: Parent directory for temporary maps when streaming is enabled;
+            None uses the system temporary directory. Files are cleaned up on exit.
 
     Returns:
         ``{"outputs": {"mask": path}, "details": {...}}``; ``details`` is described
@@ -368,6 +377,8 @@ def run(
         ModelExecutionError: Inference failed or produced a mask off the source grid
             or with labels outside the bundle's set.
     """
+    if type(streaming) is not bool:
+        raise InvalidInputError("streaming must be boolean")
     values = {
         "arterial_path": arterial_path,
         "venous_path": venous_path,
@@ -419,16 +430,29 @@ def run(
     else:
         image, inverse = preprocess_brats(volumes), None
     prepared = time.perf_counter()
-    logits, runtime = predict(network, image, config)
-    predicted = time.perf_counter()
-    if bundle == "renalStructures_CECT_segmentation":
-        assert inverse is not None
-        mask = postprocess_renal(logits[0], inverse)
-    elif bundle in _WHOLE_BODY_PIXDIM:
-        assert inverse is not None
-        mask = postprocess_wholebody(logits[0], inverse)
+    if streaming:
+        from medmlx_core.volume_workspace import VolumeWorkspace
+
+        context = VolumeWorkspace(scratch_dir)
     else:
-        mask = postprocess_brats(logits[0])
+        context = nullcontext(None)
+    with context as workspace:
+        logits, runtime = (
+            predict(network, image, config)
+            if workspace is None
+            else predict(network, image, config, workspace=workspace)
+        )
+        predicted = time.perf_counter()
+        if workspace is not None:
+            mask = postprocess_streaming(logits[0], inverse, bundle, workspace)
+        elif bundle == "renalStructures_CECT_segmentation":
+            assert inverse is not None
+            mask = postprocess_renal(logits[0], inverse)
+        elif bundle in _WHOLE_BODY_PIXDIM:
+            assert inverse is not None
+            mask = postprocess_wholebody(logits[0], inverse)
+        else:
+            mask = postprocess_brats(logits[0])
     _validate_mask(mask, volumes[0], config)
     output_root.mkdir(parents=True, exist_ok=True)
     write_atomic(mask_path, mask, volumes[0].affine, config.output_dtype)
@@ -440,6 +464,7 @@ def run(
             "weights_revision": weights_revision(weights.parent),
             "bundle": bundle,
             "arithmetic_profile": arithmetic_profile,
+            "streaming": streaming,
             "labels": {str(value): name for value, name in config.labels.items()},
             "labels_present": present,
             "input_shape": list(volumes[0].shape),
@@ -482,9 +507,21 @@ def brats(
     ``segresnet-brats.safetensors``; an explicit safetensors file is also accepted.
     Automatic downloads use ``MedMLX/segresnet-brats-mlx``. All arithmetic and
     preprocessing are delegated unchanged to ``run``.
+
+    Args:
+        t1c_path: Contrast-enhanced T1 MRI NIfTI on the shared input grid.
+        t1_path: T1 MRI NIfTI on the T1c grid.
+        t2_path: T2 MRI NIfTI on the T1c grid.
+        flair_path: FLAIR MRI NIfTI on the T1c grid.
+        output_dir: Directory for the brain-tumor label map.
+        weights_path: Optional folder containing ``segresnet-brats.safetensors``;
+            omit to download and reuse the BraTS bundle. Explicit files remain accepted.
+        overwrite: Replace an existing output mask when true.
+        arithmetic_profile: ``torch-order-fp32`` (default) or ``compensated-fp32``;
+            ``compensated-fp48`` requires the renal bundle.
     """
     folder = (
-        weights_dir("segresnet-brats-mlx")
+        weights_dir("segresnet-brats")
         if weights_path is None else Path(weights_path).expanduser()
     )
     checkpoint = folder / "segresnet-brats.safetensors" if folder.is_dir() else folder
@@ -512,9 +549,20 @@ def renal_cect(
     this bundle; an explicit safetensors file is also accepted. Automatic downloads
     use ``MedMLX/segresnet-renal-cect-mlx``. Arithmetic and preprocessing are
     delegated unchanged to ``run``.
+
+    Args:
+        arterial_path: Arterial-phase CT NIfTI on the shared input grid.
+        venous_path: Venous-phase CT NIfTI on the arterial grid.
+        excretory_path: Excretory-phase CT NIfTI on the arterial grid.
+        output_dir: Directory for the renal label map.
+        weights_path: Optional folder containing ``segresnet-renal-cect.safetensors``;
+            omit to download and reuse the renal bundle. Explicit files remain accepted.
+        overwrite: Replace an existing output mask when true.
+        arithmetic_profile: ``torch-order-fp32`` (default), ``compensated-fp32``
+            or ``compensated-fp48``; see ``run`` for their arithmetic contracts.
     """
     folder = (
-        weights_dir("segresnet-renal-cect-mlx")
+        weights_dir("segresnet-renal-cect")
         if weights_path is None else Path(weights_path).expanduser()
     )
     checkpoint = folder / "segresnet-renal-cect.safetensors" if folder.is_dir() else folder
@@ -636,6 +684,48 @@ def postprocess_wholebody(logits: Float32Array, inverse: SpatialInverse) -> Floa
     return values[0]
 
 
+def postprocess_streaming(
+    logits: Float32Array,
+    inverse: SpatialInverse | None,
+    bundle: str,
+    workspace: VolumeWorkspace,
+) -> MaskArray:
+    """Retain each bundle's activation/restoration/label order with bounded class storage."""
+    from medmlx_core.volume_workspace import restored_argmax
+
+    if bundle in _WHOLE_BODY_PIXDIM:
+        assert inverse is not None
+        probabilities = workspace.map_channels(logits, transforms.softmax_channel)
+        labels = restored_argmax(probabilities, lambda channel: channel)[None]
+        values, affine = transforms.invert_spacing(
+            labels, inverse.affine, inverse.resampling, "nearest"
+        )
+        return transforms.invert_orientation(values, affine, inverse.orientation)[0][0]
+    if bundle == "renalStructures_CECT_segmentation":
+        assert inverse is not None
+        labels: Float32Array | None = None
+        for channel in range(1, logits.shape[0]):
+            values, affine = transforms.invert_spacing(
+                logits[channel : channel + 1], inverse.affine, inverse.resampling
+            )
+            values, _ = transforms.invert_orientation(values, affine, inverse.orientation)
+            active = transforms.threshold(transforms.sigmoid(values), 0.5)[0] != 0
+            if labels is None:
+                labels = np.zeros(active.shape, dtype=np.float32)
+            labels[(labels == 0) & active] = channel
+        assert labels is not None
+        return labels
+    if bundle != "brats_mri_segmentation":
+        raise InvalidInputError(f"Unsupported streaming SegResNet bundle: {bundle}")
+    flat = logits.reshape(logits.shape[0], -1)
+    output = np.empty(flat.shape[1], dtype=np.int64)
+    for start in range(0, flat.shape[1], 65536):
+        stop = min(start + 65536, flat.shape[1])
+        binary = transforms.threshold(transforms.sigmoid(flat[:, start:stop]), 0.5)
+        output[start:stop] = transforms.brats_labels(binary)[0]
+    return output.reshape(logits.shape[1:])
+
+
 def _weights_file(weights_path: str) -> Path:
     """Validate a safetensors file path while retaining any Hub-cache symlink."""
     # MLX identifies the file format from the supplied name. Hub cache paths are
@@ -678,9 +768,8 @@ def load_inputs(roles: list[tuple[str, str | None]]) -> list[Volume]:
 def load_volume(path: Path, role: str) -> Volume:
     """Read 3D voxels using the affine selected by the upstream nibabel reader.
 
-    Source voxels are checked for finiteness before conversion to float32; values
-    outside the float32 range can overflow during that conversion. Spatial units
-    and affine invertibility are not checked here.
+    Voxels must be real, finite and representable as float32. Spatial units must be
+    mm or unspecified (assumed mm), and the corrected affine must be invertible.
     """
     from segresnet_mlx._nifti import ImageFileError, nib
 
@@ -690,15 +779,28 @@ def load_volume(path: Path, role: str) -> Volume:
         image = nib.load(path.as_posix())
     except (OSError, ImageFileError) as error:
         raise InvalidInputError(f"Invalid {role} NIfTI: {path}") from error
+    try:
+        spatial_units, _ = image.header.get_xyzt_units()
+    except (AttributeError, KeyError, ValueError) as error:
+        raise InvalidInputError(f"{role} has invalid NIfTI spatial units") from error
+    if spatial_units not in ("mm", "unknown"):
+        raise InvalidInputError(f"{role} spatial units must be mm or unspecified")
     raw = np.asanyarray(image.dataobj)
     if raw.ndim != 3:
         raise InvalidInputError(f"{role} must be a 3D NIfTI; got shape {tuple(raw.shape)}")
+    if np.iscomplexobj(raw):
+        raise InvalidInputError(f"{role} must contain real-valued voxels")
     if not np.all(np.isfinite(raw)):
         raise InvalidInputError(f"{role} contains non-finite voxels")
+    limit = np.finfo(np.float32).max
+    if np.any(raw > limit) or np.any(raw < -limit):
+        raise InvalidInputError(f"{role} contains voxels outside the finite float32 range")
     affine = np.asarray(image.affine, dtype=np.float64)
     if affine.shape != (4, 4) or not np.all(np.isfinite(affine)):
         raise InvalidInputError(f"{role} affine must be a finite 4x4 matrix")
     data, reader_affine = transforms.image_arrays(image)
+    if not np.all(np.isfinite(reader_affine)) or np.linalg.matrix_rank(reader_affine[:3, :3]) < 3:
+        raise InvalidInputError(f"{role} affine must define an invertible 3D grid")
     shape = (int(raw.shape[0]), int(raw.shape[1]), int(raw.shape[2]))
     return Volume(path=path, shape=shape, affine=reader_affine, data=data)
 
@@ -723,7 +825,11 @@ def _load_network(
 
 
 def predict(
-    network: SegResNet, image: Float32Array, config: Bundle
+    network: SegResNet,
+    image: Float32Array,
+    config: Bundle,
+    *,
+    workspace: VolumeWorkspace | None = None,
 ) -> tuple[Float32Array, Runtime]:
     """Run the bundle's ``SlidingWindowInferer`` settings around the MLX network.
 
@@ -741,9 +847,9 @@ def predict(
     try:
         host = np.ascontiguousarray(image, dtype=np.float32)
         logits = (
-            predict_wide_logits(network, host, window, mx=mx)
+            predict_wide_logits(network, host, window, mx=mx, workspace=workspace)
             if network.arithmetic_profile == "compensated-fp48"
-            else BundlePredictor(network, mx=mx).predict_logits(host, window)
+            else BundlePredictor(network, mx=mx).predict_logits(host, window, workspace=workspace)
         )
     except (RuntimeError, ValueError) as error:
         raise ModelExecutionError(f"SegResNet MLX inference failed: {error}") from error

{segresnet_mlx-0.3.0 → segresnet_mlx-0.3.1}/src/segresnet_mlx/predictor.py

@@ -13,8 +13,9 @@ from segresnet_mlx._runtime import import_mlx
 
 if TYPE_CHECKING:
     from ignite.engine import Engine, Events
+    from medmlx_core.volume_workspace import VolumeWorkspace
 
-    from segresnet_mlx._types import MLX, Array, CoreSlidingWindow
+    from segresnet_mlx._types import MLX, Array, CoreSlidingWindow, WorkspaceOptions
     from segresnet_mlx.network import SegResNet
 
 import numpy as np
@@ -169,7 +170,11 @@ class BundlePredictor:
         return cast(Float32Array, host)
 
     def predict_logits(
-        self, values: HostArray | Array, configuration: SlidingWindowConfig
+        self,
+        values: HostArray | Array,
+        configuration: SlidingWindowConfig,
+        *,
+        workspace: VolumeWorkspace | None = None,
     ) -> Float32Array:
         """Stitch preprocessed ``(N, C, D, H, W)`` float32 data into full-volume logits.
 
@@ -205,6 +210,7 @@ class BundlePredictor:
                 # MLX reduction kernels depend on layout. Match MONAI's sliced
                 # batch-one input without a second padded volume or patch copy.
                 preserve_window_strides=True,
+                workspace=workspace,
             )
         except BaseException as error:
             if self.telemetry is not None:
@@ -233,7 +239,12 @@ def predict_logits(
 
 
 def predict_wide_logits(
-    network: SegResNet, values: Float32Array, configuration: SlidingWindowConfig, *, mx: MLX
+    network: SegResNet,
+    values: Float32Array,
+    configuration: SlidingWindowConfig,
+    *,
+    mx: MLX,
+    workspace: VolumeWorkspace | None = None,
 ) -> Float32Array:
     """Carry paired Metal logits into FP64 host blending, rounding only at export.
 
@@ -258,6 +269,7 @@ def predict_wide_logits(
         del result
         return output
 
+    workspace_options: WorkspaceOptions = {} if workspace is None else {"workspace": workspace}
     try:
         logits = sliding_window_inference(
             values.astype(np.float64),
@@ -268,8 +280,9 @@ def predict_wide_logits(
             mode=configuration.mode,
             padding_mode=configuration.padding_mode,
             predictor_input="numpy",
+            **workspace_options,
         )
-        return logits.astype(np.float32)
+        return logits.astype(np.float32) if workspace is None else workspace.float32(logits)
     finally:
         # Reuse same-shaped Metal allocations across windows, including when
         # paired arithmetic requires a wider host accumulator.

{segresnet_mlx-0.3.0 → segresnet_mlx-0.3.1}/src/segresnet_mlx/telemetry.py

@@ -20,6 +20,8 @@ import numpy as np
 
 
 class EventEngine(Protocol):
+    """Structural type for the Ignite engine that receives telemetry handlers."""
+
     def add_event_handler(self, event_name: Events, handler: Callable[..., None]) -> object:
         """Register a handler for an Ignite lifecycle event."""
         ...

{segresnet_mlx-0.3.0 → segresnet_mlx-0.3.1}/src/segresnet_mlx/transforms.py

@@ -11,7 +11,7 @@ as ``Spacing`` does and returned as float32.
 
 Image arrays are channel-first float32 ``(C, *spatial)`` and affines are 4x4 float64
 voxel-to-world matrices as stored in the NIfTI header. Bundle ``pixdim`` targets
-are in mm; these functions do not inspect or convert the input's spatial units.
+are in mm; the runner validates spatial units, and these functions assume mm.
 """
 
 from __future__ import annotations
@@ -333,7 +333,7 @@ def brats_labels(binary: Float32Array) -> npt.NDArray[np.int64]:
 
 
 def write_nifti(path: Path, data: HostArray, affine: Float64Array, dtype: npt.DTypeLike) -> None:
-    """Cast and write a NIfTI with sform and qform code 1.
+    """Cast and write a millimetre NIfTI with sform and qform code 1.
 
     The sform stores the affine within NIfTI precision. The qform cannot
     represent shear, which NiBabel removes when encoding that form.
@@ -341,6 +341,7 @@ def write_nifti(path: Path, data: HostArray, affine: Float64Array, dtype: npt.DT
     array = np.asarray(data).astype(dtype, copy=False)
     matrix = to_affine_nd(3, affine)
     image = nib.Nifti1Image(array, matrix)
+    image.header.set_xyzt_units("mm")
     image.set_sform(matrix, code=1)
     image.set_qform(matrix, code=1)
     nib.save(image, str(path))